00:00:02.222 --> 00:00:07.280
GRETCHEN HUIZINGA: Welcome to Abstracts, 
a Microsoft Research Podcast that puts the  

00:00:07.280 --> 00:00:15.920
spotlight on world-class research in brief. 
I’m Gretchen Huizinga. In this series,  

00:00:15.920 --> 00:00:18.960
members of the research community 
at Microsoft give us a quick  

00:00:18.960 --> 00:00:25.280
snapshot – or a podcast abstract – 
of their new and noteworthy papers.

00:00:25.280 --> 00:00:31.280
Today I'm talking to two researchers, Hongxia 
Hao, a senior researcher at Microsoft Research  

00:00:31.280 --> 00:00:37.040
AI for Science, and Bing Lv, an associate 
professor in physics at the University of  

00:00:37.040 --> 00:00:42.400
Texas at Dallas. Hongxia and Bing are 
co-authors of a paper called Probing  

00:00:42.400 --> 00:00:47.600
the Limit of Heat Transfer in Inorganic 
Crystals with Deep Learning. I'm excited  

00:00:47.600 --> 00:00:52.787
to learn more about this! Hongxia and Bing, 
it's great to have you both on Abstracts!

00:00:52.787 --> 00:00:53.991
HONGXIA HAO: Nice to be here.

00:00:53.991 --> 00:00:55.280
BING LV: Nice to be here, too.

00:00:55.280 --> 00:00:59.440
HUIZINGA: So Hongxia, let's start with 
you and a brief overview of this paper.  

00:00:59.440 --> 00:01:03.520
In just a few sentences. Tell us 
about the problem your research  

00:01:03.520 --> 00:01:07.040
addresses and more importantly, 
why we should care about it.

00:01:07.040 --> 00:01:12.720
HAO: Let me start with a very simple yet profound 
question. What's the fastest the heat can travel  

00:01:12.720 --> 00:01:18.960
through a solid material? This is not just an 
academic curiosity, but it's a question that  

00:01:18.960 --> 00:01:24.640
touched the bottom of how we build technologies 
around us. So from the moment when you tap your  

00:01:24.640 --> 00:01:29.440
smartphone, and the moment where the laptop 
is turned on and functioning, heat is always  

00:01:29.440 --> 00:01:35.680
flowing. So we're trying to answer the question 
of a century-old mystery of the upper limit of  

00:01:35.680 --> 00:01:41.040
heat transfer in solids. So we care about this 
not just because it's a fundamental problem in  

00:01:41.040 --> 00:01:47.120
physics and material science, but because solving 
it could really rewrite the rulebook for designing  

00:01:47.120 --> 00:01:54.960
high-efficiency electronics and sustainable 
energy, etc. And nowadays, with very cutting-edge  

00:01:54.960 --> 00:02:01.760
nanometer chips or very fancy technologies, we are 
packing more computing power into smaller space,  

00:02:01.760 --> 00:02:08.720
but the faster and denser we build, the harder 
it becomes to remove the heat. So in many ways,  

00:02:08.720 --> 00:02:15.600
thermal bottlenecks, not just transistor density, 
are now the ceiling of the Moore’s Law. And also  

00:02:15.600 --> 00:02:23.040
the stakes are very enormous. We really wish to 
bring more thermal solutions by finding more high  

00:02:23.040 --> 00:02:29.360
thermal conductor choices from the perspective 
of materials discovery with the help of AI.

00:02:32.160 --> 00:02:47.200
LV: So I think one of the biggest things 
as Hongxia said, right? Thermal solutions  

00:02:47.200 --> 00:02:52.800
will become, eventually become, a bottleneck 
for all type of heterogeneous integration of  

00:02:52.800 --> 00:03:01.200
the materials. So from this perspective, so how 
people actually have been finding out previously,  

00:03:01.200 --> 00:03:06.880
all the thermal was the last solution 
to solve. But now people actually more  

00:03:06.880 --> 00:03:11.600
and more realize all these things 
have to be upfront. This co-design,  

00:03:11.600 --> 00:03:17.520
all these things become very important. 
So I think what we are doing right now,  

00:03:17.520 --> 00:03:25.360
integrated with AI, helping to identify the large 
space of the materials, identify fundamentally  

00:03:25.360 --> 00:03:30.385
what will be the limit of this material, 
will become very important for the society.

00:03:30.385 --> 00:03:33.720
HUIZINGA: Hmm. Yeah. Hongxia, did 
you have anything to add to that?

00:03:33.720 --> 00:03:40.080
HAO: Yes, so previously many people are working 
on exploring these material science questions  

00:03:40.080 --> 00:03:46.160
through experimental tradition and the 
past few decades people see a new trend  

00:03:46.160 --> 00:03:52.320
using computational materials discovery. Like 
for example, we do the fundamental solving of  

00:03:52.320 --> 00:03:56.880
the Schrödinger equation using Density Functional 
Theory [DFT]. Actually, this brings us a lot of  

00:03:56.880 --> 00:04:02.720
opportunities. The question here is, as the 
theory is getting more and more developed,  

00:04:02.720 --> 00:04:09.120
it’s too expensive for us to make it very large 
scale and to study tons of materials. Think about  

00:04:09.120 --> 00:04:15.520
this. The bottleneck here, now, is not just 
about having a very good theory, it's about  

00:04:15.520 --> 00:04:22.600
the scale. So, there is where AI, specifically 
now we are using deep learning, comes into play.

00:04:22.600 --> 00:04:27.040
HUIZINGA: Well, Hongxia, let's stay with 
you for a minute and talk about methodology.  

00:04:27.040 --> 00:04:31.720
How did you do this research and what 
was the methodology you employed?

00:04:31.720 --> 00:04:37.520
HAO: So here we, for this question, 
we built a pipeline that spans the AI,  

00:04:37.520 --> 00:04:42.960
the quantum mechanics, and computational 
brute-force with a blend of efficiency  

00:04:42.960 --> 00:04:49.520
and accuracy. It begins with generating an 
enormous chemical and structure design space  

00:04:49.520 --> 00:04:56.000
because this is inspired by Slack’s principle. We 
focus first on simple crystals, and there are the  

00:04:56.000 --> 00:05:01.840
systems most likely to have low and harmonious 
state, fewer phononic scattering events,  

00:05:01.840 --> 00:05:06.720
and therefore potentially have high thermal 
conductivities. But we didn't stop here. We  

00:05:06.720 --> 00:05:13.280
also included a huge pool of more complex and 
higher energy structures to ensure diversity  

00:05:13.280 --> 00:05:20.320
and avoid bias. And for each candidate, we first 
run like a structure relaxation using MatterSim,  

00:05:20.320 --> 00:05:26.560
which is a deep learning foundational model 
for material science for us to characterize  

00:05:26.560 --> 00:05:34.240
the properties of materials. And we use that 
screen for dynamic stability. And now it's about  

00:05:34.240 --> 00:05:41.760
200K structures past this filter. And then came 
another real challenge: calculating the thermal  

00:05:41.760 --> 00:05:47.760
conductivity. We try to solve this problem 
using the Boltzmann transport equation and  

00:05:47.760 --> 00:05:54.720
the three-phonon scattering process. The twist 
here is all of this was not done by traditional  

00:05:54.720 --> 00:06:01.360
DFT solvers, but with our deep learning model, the 
MatterSim. It's trained to predict energy, force,  

00:06:01.360 --> 00:06:07.760
and stress. And we can get second- and third-order 
interatomic force constants directly from here,  

00:06:07.760 --> 00:06:14.800
which can guarantee the accuracy of the solution. 
And finally, to validate the model's predictions,  

00:06:14.800 --> 00:06:20.560
we performed full DFT-based calculations 
on the top candidates that we found,  

00:06:20.560 --> 00:06:28.560
some of which even include higher-order scattering 
mechanism, electron phonon coupling effect, etc.  

00:06:28.560 --> 00:06:35.680
And this rigorous validation gave us confidence in 
the speed and accuracy trade-offs and revealed a  

00:06:35.680 --> 00:06:41.960
spectrum of materials that had either previously 
been overlooked or were never before conceived.

00:06:41.960 --> 00:06:46.160
HUIZINGA: So Bing, let's talk 
about your research findings.  

00:06:46.160 --> 00:06:49.600
How did things work out for you on 
this project and what did you find?

00:06:49.600 --> 00:06:56.320
LV: I think one of the biggest things 
for this paper is it creates a very  

00:06:56.320 --> 00:07:02.240
large material base. Basically, you can say 
it's a smart database which eventually will  

00:07:02.240 --> 00:07:06.800
be made accessible to the public. I think 
that's a big achievement because people who  

00:07:06.800 --> 00:07:11.920
actually if they have to look into it, they 
actually can go search Microsoft database,  

00:07:11.920 --> 00:07:16.800
finding out, oh, this material does have this 
type of thermal properties. This is actually,  

00:07:16.800 --> 00:07:27.200
this database can send about 230,000 materials. 
And one of the things we confirm is the highest  

00:07:27.200 --> 00:07:34.480
thermal conductivity material based on all the 
wisdom of Slack criteria, predicted diamond would  

00:07:34.480 --> 00:07:41.280
have the highest thermal conductivity. We more 
or less really very solidly prove diamond, at  

00:07:41.280 --> 00:07:46.880
this stage, will remain with the highest thermal 
conductivity. We have a lot of new materials,  

00:07:46.880 --> 00:07:52.560
exotic materials, which some of them, Hongxia can 
elaborate a little bit more. So, which having all  

00:07:52.560 --> 00:07:58.160
this very exotic combination of properties, 
thermal with other properties, which could  

00:07:58.160 --> 00:08:03.680
actually provide a new insight for new physics 
development, new material development, and a  

00:08:03.680 --> 00:08:09.280
new device perspective. All of this combined will 
have actually a very profound impact to society.

00:08:09.280 --> 00:08:13.680
HUIZINGA: Yeah, Hongxia, go a little deeper on 
that because that was an interesting part of the  

00:08:13.680 --> 00:08:18.960
paper when you talked about diamond still being 
the sort of “gold standard,” to mix metaphors!  

00:08:18.960 --> 00:08:23.920
But you've also found some other materials 
that are remarkable compared to silicon.

00:08:23.920 --> 00:08:29.920
HAO: Yeah, yeah. Among this search space, 
even though we didn't find like something  

00:08:29.920 --> 00:08:35.360
that's higher than diamonds, but we do discover 
more than like twenty new materials with thermal  

00:08:35.360 --> 00:08:40.960
conductivity exceeding that of silicon. 
And silicon is something like a benchmark  

00:08:40.960 --> 00:08:46.480
for criteria that we think we want to 
compare with because it's a backbone of  

00:08:46.480 --> 00:08:52.693
modern electronics. More interestingly, 
I think, is the manganese vanadium.  

00:08:53.600 --> 00:08:59.600
It shows some very interesting and surprising 
phenomena. Like it's a metallic compound,  

00:08:59.600 --> 00:09:05.920
but with very high lattice thermal connectivity. 
And this is the first time discovered by, like,  

00:09:05.920 --> 00:09:13.840
through our search pattern, and it’s something 
that cannot be easily discovered without the hope  

00:09:13.840 --> 00:09:20.800
with AI. And right now, think Bing can explain 
more on this, and show some interesting results.

00:09:20.800 --> 00:09:22.720
HUIZINGA: Yeah, go ahead Bing.

00:09:22.720 --> 00:09:29.440
LV: So this is actually very surprising to me 
as an experimentalist because of when Hongxia  

00:09:29.440 --> 00:09:38.160
presented their theory work to me, this material, 
magnesium vanadium, it's discovered back in 1938,  

00:09:38.160 --> 00:09:45.520
almost 100 years ago, but there's no more 
than twenty papers talking about this! A  

00:09:45.520 --> 00:09:50.720
lot of them was on theory, okay, not even 
on experimental part. We actually did quite  

00:09:50.720 --> 00:09:56.640
a bit of work on this. We actually are in the 
process; will characterize this and then moving  

00:09:56.640 --> 00:10:00.720
forward even for the thermal conductivity 
measurements. So that will be hopefully,  

00:10:00.720 --> 00:10:05.520
will be adding to the value of these things, 
showing you, Hey, AI does help to predict the  

00:10:05.520 --> 00:10:10.360
materials could really generate the new materials 
with very good high thermal conductivity.

00:10:10.360 --> 00:10:14.160
HUIZINGA: Yeah, so Bing, stay with you for 
a minute. I want you to talk about some  

00:10:14.160 --> 00:10:19.520
kind of real-world applications of this. 
I know you alluded to a couple of things,  

00:10:19.520 --> 00:10:23.200
but how is this work significant in that respect,  

00:10:23.200 --> 00:10:28.236
and who might be most excited about 
it, aside from the two of you? [LAUGHS]

00:10:28.236 --> 00:10:35.360
LV: So I think as I mentioned before, the first 
thing is this database. I believe that's the  

00:10:35.360 --> 00:10:43.280
first ever large material database regarding to 
the thermal conductivity. And it has, as I said,  

00:10:43.280 --> 00:10:51.360
230,000 materials with AI-predicted thermal 
connectivity. This will provide not only  

00:10:51.360 --> 00:10:57.680
science but engineering with a vastly expanding 
catalog of candidate materials for the future  

00:10:57.680 --> 00:11:02.960
roadmap of integration, material integration, 
and all these bottlenecks we are talking about,  

00:11:02.960 --> 00:11:08.400
the thermal solution for the semiconductors or 
for even beyond the semiconductor integration,  

00:11:08.400 --> 00:11:13.040
people actually can have a database 
to looking for. So these things,  

00:11:13.040 --> 00:11:17.760
it will become very important, and 
I believe over a long time it will  

00:11:17.760 --> 00:11:22.840
generate a very long impact for the research 
community, for the society development.

00:11:22.840 --> 00:11:26.960
HUIZINGA: Yeah. Hongxia, did you 
have anything to add to that one too?

00:11:26.960 --> 00:11:34.480
HAO: Yeah, so this study reshapes how we think 
about limits. I like the sentence that the only  

00:11:34.480 --> 00:11:40.880
way to discover the limits of possible is to go 
beyond them into the impossible. In this case,  

00:11:40.880 --> 00:11:46.800
we tried, but we didn't break the diamond 
limit. But we proved it even more rigorously  

00:11:46.800 --> 00:11:52.640
than ever before. In doing so, we also 
uncovered some uncharted peaks in the  

00:11:52.640 --> 00:11:58.000
thermal conductivity landscape. This would 
not happen without new AI capabilities for  

00:11:58.000 --> 00:12:03.040
material science. I think in the long run, 
I believe researchers could benefit from  

00:12:03.040 --> 00:12:08.800
using this AI design and shift their way 
on how to do materials research with AI.

00:12:08.800 --> 00:12:13.040
HUIZINGA: Yeah, it'll be interesting to 
see if anyone ever does break the diamond  

00:12:13.040 --> 00:12:16.300
limit with the new tools that are available, but…

00:12:16.300 --> 00:12:16.310
HAO: Yeah!

00:12:16.310 --> 00:12:19.920
HUIZINGA: So this is the part of the 
abstracts podcast where I like to ask  

00:12:19.920 --> 00:12:26.400
for sort of a golden nugget, a one sentence 
takeaway that listeners might get from this  

00:12:26.400 --> 00:12:31.600
paper. If you had one Hongxia, what would it 
be? And then I'll ask Bing to maybe give his.

00:12:31.600 --> 00:12:38.880
HAO: Yes. AI is no longer just a tool. 
It's becoming a critical partner for us in  

00:12:38.880 --> 00:12:45.840
scientific discovery. So our work proved that the 
large-scale data-driven science can now approach  

00:12:45.840 --> 00:12:50.880
long-standing and fundamental questions 
with very fresh eyes. When trained well,  

00:12:50.880 --> 00:12:56.800
and guided with physical intuition, models 
like MatterSim can really realize a full  

00:12:56.800 --> 00:13:03.120
in-silico characterization for materials and 
don't just simulate some known materials,  

00:13:03.120 --> 00:13:10.400
but really trying to imagine what nature hasn't 
yet revealed. Our work points to a path forward,  

00:13:10.400 --> 00:13:13.840
not just incrementally better 
materials, but entirely new  

00:13:13.840 --> 00:13:19.600
class of high-performance compounds where 
we could never have guessed without AI.

00:13:19.600 --> 00:13:22.960
HUIZINGA: Yeah. Bing, what's your one takeaway?

00:13:22.960 --> 00:13:29.120
LV: I think I want to add a few things on 
top of Hongxia’s comments because I think  

00:13:29.120 --> 00:13:36.400
Hongxia has very good critical words I would 
like to emphasize. When we train the AI well,  

00:13:36.400 --> 00:13:45.520
if we guide the AI well, it could be very useful 
to become our partner. So I think all in all,  

00:13:45.520 --> 00:13:52.160
our human being’s intellectual merit here is still 
going to play a significantly important role,  

00:13:52.160 --> 00:13:57.760
okay? We are generating this AI, we should 
really train the AI, we should be using our  

00:13:57.760 --> 00:14:06.880
human being intellectual merit to guide them to 
be useful for our human being society advancement.  

00:14:06.880 --> 00:14:13.440
Now with all these AI tools, I think it's a very 
golden time right now. Experimentalists could  

00:14:13.440 --> 00:14:18.400
work very closely with like Hongxia, who’s a good 
theorist who has very good intellectual merits,  

00:14:18.400 --> 00:14:25.600
and then we actually now incorporate with 
AI, then combine all pieces together,  

00:14:25.600 --> 00:14:31.120
hopefully we’re really able to accelerating 
material discovery in a much faster pace than  

00:14:31.120 --> 00:14:35.360
ever which the whole society will 
eventually get a benefit from it.

00:14:35.360 --> 00:14:41.040
HUIZINGA: Yeah. Well, as we close, Bing, 
I want you to go a little further and talk  

00:14:41.040 --> 00:14:46.160
about what's next then, research wise. What are 
the open questions or outstanding challenges  

00:14:46.160 --> 00:14:51.200
that remain in this field and what's on 
your research agenda to address them?

00:14:51.200 --> 00:15:00.080
LV: So first of all, I think this paper is 
addressing primarily on these crystalline ordered  

00:15:00.080 --> 00:15:05.680
inorganic bulk materials. And also with the 
condition we are targeting at ambient pressure,  

00:15:05.680 --> 00:15:10.560
room temperature, because that's normally 
how the instrument is working, right? But  

00:15:11.520 --> 00:15:17.120
what if under extreme conditions? We want to go to 
space, right? There we’ll have extreme conditions,  

00:15:17.120 --> 00:15:22.640
some very… sometimes very cold, sometimes very 
hot. We have some places with extremely probably  

00:15:22.640 --> 00:15:29.280
quite high pressure. Or we have some conditions 
that are highly radioactive. So under that  

00:15:29.280 --> 00:15:34.480
condition, there’s going to be a new database 
could be emerged. Can we do something beyond  

00:15:34.480 --> 00:15:40.720
that? Another good important thing is we are 
targeting this paper on high thermal conductivity.  

00:15:40.720 --> 00:15:46.960
What about extremely low thermal conductivity? 
Those will actually bring a very good challenge  

00:15:46.960 --> 00:15:51.600
for theorists and also the machine learning 
approach. I think that's something Hongxia  

00:15:51.600 --> 00:15:55.440
probably is very excited to work on in that 
direction. I know since she’s ambitious,  

00:15:55.440 --> 00:16:00.200
she wants to do something more than 
beyond what we actually achieved so far.

00:16:00.200 --> 00:16:06.360
HUIZINGA: Yeah, so Hongxia, how would you 
encapsulate what your dream research is next?

00:16:06.360 --> 00:16:12.400
HAO: Yeah, so I think besides all of 
these exciting research directions,  

00:16:12.400 --> 00:16:18.720
on my end, another direction is perhaps 
kind of exciting is we want to move from  

00:16:18.720 --> 00:16:23.920
search to design. So right now we are 
kind of good at asking like what exists  

00:16:23.920 --> 00:16:29.440
by just doing a forward prediction and 
brute force. But with generative AI,  

00:16:29.440 --> 00:16:36.320
we can start asking what should exist? In the 
future, we can have an incorporation between  

00:16:36.320 --> 00:16:43.440
forward prediction and backwards generative 
design to really tackle questions. If you have  

00:16:43.440 --> 00:16:49.200
materials like you want to have desired like 
properties, how would you design the problems?

00:16:49.200 --> 00:16:54.800
HUIZINGA: Well, it sounds like there's 
a full plate of research agenda goodness  

00:16:54.800 --> 00:17:01.440
going forward in this field, both with human 
brains and AI. So, Hongxia Hao and Bing Lv,  

00:17:01.440 --> 00:17:05.760
thanks for joining us today. And to 
our listeners, thanks for tuning in.  

00:17:05.760 --> 00:17:12.320
If you want to read this paper, you 
can find a link at aka.ms/Abstracts,  

00:17:12.320 --> 00:17:32.320
or you can read a pre-print of it on 
arXiv. See you next time on Abstracts!